A Physics-Guided Neural Network-Based Full Waveform Inversion Method and its Forward Modeling Study

Seismic data are essential for subsurface parameter inversion, with their resolution and completeness directly impacting reservoir prediction and geological interpretation. While conventional full waveform inversion (FWI) offers high-resolution imaging, its reliance on accurate initial models, susceptibility to cycle skipping, and high computational cost hinder its practical application. This study proposes a Physics-Guided Neural Network FWI (PGNN-FWI) approach that embeds the wave equation as a physical constraint within a neural network framework to improve inversion accuracy and robustness. Comparative experiments show that PGNN-FWI achieves superior performance over traditional FWI in terms of reconstruction accuracy, convergence rate, and noise tolerance. Forward modeling using the 2D acoustic wave equation further confirms that velocity models from PGNN-FWI yield synthetic seismic data with more detailed and consistent waveforms. These results demonstrate the effectiveness of PGNN-FWI and highlight its potential in seismic imaging and interpretation.